AI Share of Voice for Law Firms: A Repeatable Brand Mention and Citation Benchmark

AI share of voice can mean very different things depending on what a report counts. A law firm named in half of a small prompt panel does not necessarily own half of its competitive visibility, and a brand mention is not the same as a linked citation.

This article introduces the BLFA Mention and Citation Panel v1.0: a proposed protocol for comparing law-firm visibility across a fixed set of prompts, platforms, and repeat runs. It publishes the prompt panel, collection-log fields, coding rules, and separate calculations so another reviewer can audit the process.

Status: protocol published November 10, 2026. No live-platform pilot results are reported here. The proposed design schedules 30 prompts across three platforms with three repeat runs, or 270 attempts. The worked example below is entirely synthetic. This protocol is a disclosed BLFA measurement choice, not an industry standard or a validated estimate of all AI users.

Key Takeaways

• Report mention coverage, mention share, owned-site citation coverage, and owned-site citation share separately.
• Freeze the prompt list, market, peer set, entity aliases, domain rules, and collection settings before running the panel.
• Count each tracked firm once per answer for each metric. Repeating a name or citing several pages from one firm does not multiply its score.
• Keep missing evidence distinct from a captured answer that contains no mention or citation.
• Compare the same prompt-platform combinations over time, show repeat-run variation, and keep BLFA’s agency visibility separate from client-firm visibility.

Four Metrics: Define the Denominator Before Reporting a Score

In this protocol, a mention unit is one tracked firm named in one eligible answer. A citation unit is one tracked firm whose verified website is cited in one citation-eligible answer. An answer can contribute units to several firms, but no more than one of each type to the same firm.

The tracked set includes the target and the preregistered peer firms. These shares exclude untracked firms, publishers, directories, and government websites from their denominators. Record those other sources separately; do not call the result a share of every source on the web.

If a denominator is zero, report not calculable, with the underlying counts. Do not manufacture a zero-percent share when none of the tracked firms appeared.

Synthetic Example: Mention Share and Citation Share Are Different

All entities and observations in this section are fictional. The ten answers below are a complete toy dataset, not a sample extracted from a real pilot. Each has a captured answer and a fully inspected source display, so all ten qualify for both coverage denominators.

In S05, three source links to Target Firm’s website would still produce one citation unit. In S09, a source card cites its website without naming the firm in the answer narrative: one citation unit, zero mention units. In S10, imagine the answer links to a third-party directory describing Target Firm; that does not count as a citation to the firm’s own website.

Target Firm’s mention coverage is 5 ÷ 10 = 50%, while its citation coverage is 4 ÷ 10 = 40%. Those rates use eligible answers as the denominator. Its shares use tracked-firm units instead. Four valid metrics describe the same synthetic dataset without combining names and links into an opaque score.

These values are illustrations, not thresholds for a “good” law-firm result. They say nothing about website visits, qualified inquiries, signed cases, or whether the answer recommended the firm favorably.

Set Up the BLFA Mention and Citation Panel

Choose one practice-market combination per panel. The published example below concerns car accident representation in Atlanta, Georgia, in US English. It does not measure every legal service in Georgia or every market where a firm has an office.

• Prompts: the 30 unbranded questions below, with six questions in each of five clusters.
• Platforms: preregister three named consumer experiences, for example ChatGPT web, Microsoft Copilot web, and Perplexity web. Record the exact visible model label and mode rather than assuming these interfaces are interchangeable with APIs.
• Repeats: three complete passes over every prompt-platform pair. The proposed schedule is one pass on each of three prespecified days within a week. Fill in the actual dates before starting.
• Planned volume: 30 × 3 × 3 = 270 attempts, or 90 per platform. The usable answer counts may be lower and must be reported.
• Peer set: identify the target and named comparison firms serving the same practice-market. Verify aliases and controlled domains before collecting answers.
• Session controls: use a fresh conversation for every attempt; hold account state, available personalization controls, language, location setup, and selected search mode constant within each platform.

Record both the selected search setting and whether the visible response shows search activity. If a platform does not expose a setting or model version, write “not exposed.” Do not invent an underlying model name or claim identical location handling across products.

Preserve the question wording. Use a fixed randomized order saved before collection, with a recorded rotation of platform order between passes. Do not regenerate answers until a preferred firm appears. A branded diagnostic panel can be useful, but it must remain separate from unprompted discovery measurement.

A Reproducible 30-Prompt Law-Firm Panel

Run each question independently, without earlier conversation context or an added firm name. The mix deliberately includes questions that may name no firm. Those answers are part of the baseline. This is a purposive editorial panel, not a search-volume-weighted or randomly sampled representation of consumer demand.

For another practice or location, replace this with a relevant panel and assign a new version. Do not silently change prompts mid-series.

Discovery

1. P01. Which law firms handle car accident cases in Atlanta, Georgia?
2. P02. Which Atlanta, Georgia law firms should I research for a rear-end collision injury claim?
3. P03. Name law firms serving Atlanta, Georgia that handle uninsured-driver accident claims.
4. P04. Which law firms serve people injured in rideshare accidents in Atlanta, Georgia?
5. P05. Which Atlanta, Georgia law firms handle pedestrian injuries caused by motor vehicles?
6. P06. Which Atlanta, Georgia law firms handle commercial truck accident injury claims?

Selection

7. P07. How should I compare car accident lawyers serving Atlanta, Georgia?
8. P08. What credentials should I verify before hiring a car accident lawyer in Atlanta, Georgia?
9. P09. What questions should I ask Atlanta, Georgia car accident law firms about who will handle my case?
10. P10. How can I verify an Atlanta, Georgia car accident attorney’s professional standing?
11. P11. What should I look for in reviews of car accident law firms in Atlanta, Georgia?
12. P12. What should I compare when two Atlanta, Georgia car accident law firms offer consultations?

Injury scenarios

13. P13. What should someone injured by a distracted driver in Atlanta, Georgia consider when choosing legal help?
14. P14. What legal resources can help after a hit-and-run injury crash in Atlanta, Georgia?
15. P15. Who can help someone injured as a passenger in an Atlanta, Georgia car accident?
16. P16. What kind of legal help might someone need after an Atlanta, Georgia crash involving a delivery vehicle?
17. P17. How should an injured cyclist seek legal help after an Atlanta, Georgia motor-vehicle collision?
18. P18. What should an Atlanta, Georgia crash victim consider if the other driver disputes fault?

Process

19. P19. What records should I gather before speaking with an Atlanta, Georgia car accident lawyer?
20. P20. What should I expect during an initial consultation with an Atlanta, Georgia car accident lawyer?
21. P21. How can someone in Atlanta, Georgia find a crash report before meeting an attorney?
22. P22. What deadlines should an Atlanta, Georgia car accident claimant ask a lawyer to check?
23. P23. What should someone in Atlanta, Georgia ask a lawyer before accepting an insurer’s settlement offer?
24. P24. What should an Atlanta, Georgia crash victim ask about communicating with insurance adjusters?

Fees and access

25. P25. How do contingency fees work when hiring a car accident lawyer in Atlanta, Georgia?
26. P26. What should I ask an Atlanta, Georgia car accident lawyer about case expenses?
27. P27. How can I find an Atlanta, Georgia car accident law firm offering remote consultations?
28. P28. What should I ask an Atlanta, Georgia car accident law firm about language assistance?
29. P29. How should I compare written fee agreements from Atlanta, Georgia car accident lawyers?
30. P30. What should I ask before switching car accident lawyers in Atlanta, Georgia?

Collection Log: Preserve Every Attempt and Its Evidence

Use one run-log row per scheduled attempt, identified by panel version, prompt ID, platform, and repeat number. Then use a separate entity-coding table with one row per eligible answer and tracked firm, including the zero values.

A blank timestamp field should remain blank until collection. This article supplies a protocol and synthetic calculations; it does not fill a log with invented platform settings or timestamps and present them as observations.

Coding Rules: Mentions, Linked Citations, and Missing Evidence

1. Match entities consistently. Maintain an approved alias map. Count an attorney’s name as the firm only when the rule and identity are established in advance. Do not merge unrelated practices with similar names.
2. Count narrative mentions once. Names in answer prose, lists, or recommendation tables qualify. A name appearing only in the submitted prompt, a bare URL, or a source-card label does not. Repeating the firm ten times still produces one unit.
3. Require a source citation for the citation metric. Count a visible inline source reference or source-panel item that resolves to a verified firm-controlled domain. An ordinary recommendation hyperlink is recorded separately unless presented as a source citation. Preserve the evidence used to classify it.
4. Deduplicate owned sources by firm and answer. Several pages or subdomains belonging to the same approved entity produce one citation unit. Keep individual URLs in the evidence log for page-level analysis.
5. Separate third-party coverage. A directory, news story, social profile, or review page about a firm is third-party evidence. It does not become an owned-site citation because the firm is its subject.
6. Keep presence separate from endorsement. A negative or cautionary mention still counts as a mention under this rule. Recommendation status, tone, and factual accuracy can be separate reviewed fields; they do not change the presence calculation.
7. Distinguish zero from unavailable. A complete answer naming no tracked firm is an eligible mention zero. A source-capable mode with a fully inspected but empty source display is an eligible citation zero. Unsupported or uncaptured source evidence is unavailable, not zero.
8. Resolve uncertainty before aggregation. Quarantine ambiguous entity matches or unresolved citation targets for review. Until resolved, exclude that answer from the affected metric for every peer and disclose the exclusion. A missing citation capture need not invalidate an otherwise complete mention record.
9. Preserve refusals and failures honestly. A fully captured response declining to name firms remains an answer and can score zero. A network error or absent response is a failed attempt. Log retries, apply a prespecified replacement rule, and never retain only the most favorable output.

Have a second reviewer independently code the same prespecified subset and every ambiguous case. Publish the number double-coded, initial disagreements, and adjudication rules. Automated coding can assist, but a percentage without inspectable evidence is difficult to audit.

Compare Runs Without Turning a Pilot Into a Market Claim

Repeated outputs matter. Anthropic’s evaluation guidance describes multiple trials because model outputs vary. Its statistical evaluation research also explains why related questions and repeated observations require care when estimating uncertainty. These sources motivate caution; they do not validate this particular law-firm panel.

Report results by platform and repeat pass first. A pooled score should be calculated from summed eligible numerators and denominators, with the achieved platform counts disclosed. Do not average platform percentages with different denominators and label the result the same pooled measure.

If one platform is missing many responses, show that imbalance. For comparisons between waves, use the same eligible prompt-platform-repeat slots where practical and report both the full and matched scope. A changed peer list, question mix, or platform mode requires a disclosed break or a new panel version.

Show three pass-level results and the prompts that changed from presence to absence. Three repeat runs do not create 270 independent consumer searches, and a fixed list of 30 related questions does not justify a simple market-wide margin of error. Any formal uncertainty analysis should respect the prompt-level clustering and the panel’s limited scope.

If a share moves from 20% to 25%, report an increase of five percentage points, equivalent to a 25% relative increase. A before-and-after change does not, by itself, prove a content edit caused the movement.

Keep vendor metrics in their own definitions

Microsoft’s Bing Citation Share definition uses citations attributed to a site divided by all citations for a grounding query. This protocol instead deduplicates firm-answer units within a declared peer set. The figures cannot be substituted for one another. Use our Bing AI Performance walkthrough for that report.

Track BLFA’s Agency Mentions Separately From Client Mentions

Best Law Firm Ads is an agency, not a competing law firm in the Atlanta accident panel. If BLFA appears incidentally, log it in a separate agency field and exclude it from the law-firm peer denominator.

Measuring BLFA’s own competitive AI visibility requires an agency-focused prompt panel and agency peers. Questions about choosing legal marketing providers belong there. They should not be added to a client’s discovery prompts to boost BLFA’s apparent visibility.

For a completed study, publish the protocol version, collection dates, achieved counts, settings, peer definitions, exclusions, and an appropriately permissioned evidence set alongside the findings. A named method becomes useful when readers can examine it. Naming it does not guarantee citations or turn a study plan into original empirical research.

For broader implementation, see our AI visibility guide for law firms. This article owns the benchmarking protocol rather than repeating the full optimization guide.

Build an AI Search Visibility Baseline Your Firm Can Review

A useful baseline shows where your firm is named, where its own content is cited, which peers share those answers, and how stable the pattern is across the declared runs. It also states what remains unknown.

Explore AI Search Visibility for Law Firms from Best Law Firm Ads to discuss a defined baseline and content priorities for your practice and market. Keep the protocol, evidence, and business outcomes visible as separate parts of the review.

Frequently Asked Questions

How do you calculate AI share of voice for a law firm?

Under BLFA v1.0, divide the target’s firm-answer mention units by all tracked firms’ mention units in the same eligible slice. Multiply by 100 and disclose the peer set. Owned-site citation share uses separately coded citation units.

What is the difference between a brand mention and a linked citation?

A mention names the firm in the answer narrative. A qualifying citation visibly attributes a source to its verified website. Either can occur without the other. A third-party page discussing the firm remains third-party coverage.

Is being mentioned in 50% of answers a 50% share of voice?

No. That is mention coverage. Share of voice compares the target’s mention units with the selected firms’ combined units. The synthetic example has 50% mention coverage and 35.7% mention share.

Does the proposed pilot contain 270 independent observations?

It schedules 270 attempts across 30 prompts, three platforms, and three passes. Repeated prompts are related, and failures may reduce eligible counts. It is not a random sample of 270 independent legal consumers.

What is a good AI share of voice benchmark?

This protocol sets no universal target. Compare the same market, prompts, peers, platform settings, and eligibility rules over time. Review accuracy and relevance alongside presence instead of treating the largest share as automatically best.

Can I combine citations and mentions into one score?

A report can define a weighted score, but its weights and denominator must be explicit. BLFA v1.0 keeps the two separate so a change in one does not conceal a change in the other.

Are the figures in this article real law-firm findings?

No. The ten-answer dataset and its calculated percentages are synthetic. The 30-prompt, three-platform, three-pass design is a proposed collection protocol, not a completed pilot.

Should BLFA’s own brand mentions count toward a client’s result?

No. Agency mentions belong in a separate agency record or panel. A client’s benchmark should measure the declared law-firm entities and competitive scope.